Documente Academic
Documente Profesional
Documente Cultură
net/publication/221757566
Article in Conference proceedings: ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine
and Biology Society. Conference · August 2011
DOI: 10.1109/IEMBS.2011.6091471 · Source: PubMed
CITATIONS READS
20 127
5 authors, including:
Some of the authors of this publication are also working on these related projects:
Morphological changes in cardiac muscle associated with progression of Diabetes Mellitus View project
All content following this page was uploaded by Herbert F Jelinek on 21 May 2014.
Abstract—Diabetic retinopathy (DR) is a complication of This paper proposes a process that does not require
diabetes, which if untreated leads to blindness. DR early preprocessing but deals with image differences of the retinal
diagnosis and treatment improve outcomes. Automated fundus directly. The approach constructs a visual dictionary to
assessment of single lesions associated with DR has been represent important features that characterize the pathology of
investigated for sometime. To improve on classification, especially
interest and uses pattern recognition tools to classify the retinal
across different ethnic groups, we present an approach using
points-of-interest and visual dictionary that contains important images into disease and non-disease.
features required to identify retinal pathology. Variation in
images of the human retina with respect to differences in II. METHODS
pigmentation and presence of diverse lesions can be analyzed
without the necessity of preprocessing and utilizing different A. Diabetic Retinopathy Images
training sets to account for ethnic differences for instance. For creating the visual words and training the detectors
(two-class classifiers), 672 non-DR images, 261 images with
I. INTRODUCTION bright lesions and 246 images with red lesions from the
Ophthalmology Dept., Federal University of São Paulo, Brazil
I NDIGENOUS populations such as the Australian
Aborigine, all have an increased incidence of diabetes
compared to the Caucasian population resident in these
were used. The images came from patients with different racial
background and were manually graded by specialists. The
countries [1]. To optimize screening, detection and treatment, images ranged from 640×640 to 1,581×1,113 pixels in
mobile screening combined with automated classification of resolution. All of these images were used for creating the
disease can be used [2]. dictionaries and for training the two-class classifiers (normal
Automated assessment of pre-proliferative diabetic vs. bright lesions and normal vs. red lesions). The aboriginal
retinopathy has been possible for some time using fluorescein- test images were obtained from the Albury Eye Clinic using a
labeled images [3]. Results for color fundus analysis Topcon camera at 1,200×1,200-pixel resolution. Retinal
identifying microaneurysms, exudates and cotton-wool spots images contained either no pathology, red lesions, or bright
as well as proliferative retinopathy have only been reported lesions.
more recently [4-6]. B. Detection of Features
To optimize automated processing of color images one has
to consider intra-image variation such as light diffusion, Every image in a collection is represented using a large
pathology, variation in fundus reflectivity and fundus number of points of interest (PoI) [9] and a local descriptor
thickness and inter-image variation (being the result of using around each PoI using the Speeded-Up Robust Features
different cameras, illumination, acquisition angle and retinal (SURF) approach is calculated [10]. This is then stored in an
pigmentation). Several methods are available including grey indexing data structure. PoIs are robustly estimated as they
world normalization, histogram equalization and histogram convey more information than other points in the image. A PoI
specification [7]. Color normalization also increases the shows repeatability after several image transformations across
discrimination in almost all of the color features [8]. different scales.
The SURF algorithm has four major stages:
Manuscript received March 26, 2011. This work was supported in part by (1) Feature point detection: this stage uses an Hessian
the Microsoft Research and Fapesp under the grants MSR-Fapesp detector approximation based on low-pass box filters (Haar
2008/54443-2 and Fapesp 2010/05647-4.
A. Rocha, T. Carvalho, S. Goldenstein, and J. Wainer are with the filters) and integral images [11] to speed up the operations.
Reasoning for Complex Data Lab., Institute of Computing, Univ. of Campinas (2) Feature point localization: The determinant of the
(Unicamp), Campinas, SP, Brazil. Contact (phone/fax: +55 19 35215854; e- Hessian for both location and scale is used. Given a point
mail: anderson.rocha@ic.unicamp.br).
H.F. Jelinek is with the Centre for Research in Complex Systems and the in an image I, the Hessian matrix in at
School of Community Health, Albury, NSW, Australia (phone: +61 2 scale is defined as follows
60519219; fax: +61 2 60519219; e-mail: hjelinek@csu.edu.au).
Hessian matrix are then interpolated in scale and image space. Input Example
to Classify PoI Finding
Description (e.g., SVM)
Aa
(3) Orientation assignment: SURF calculates the Haar- Test ing T r aining
is the scale at which the interest point was detected. Fig. 1.General pipeline of the proposed approach.
Fast filtering, in each scale , the method calculates wavelet
responses using integral images. The dominant orientation is Each image in the training generates a series of points of
estimated by calculating the sum of all responses within a interest. After finding the PoIs, the dictionary or codebook was
sliding orientation window covering an angle of . The point created, which represents distinctive regions of the images
of interest gets the orientation from the longest vector. with bright lesions as well as images tagged as normal by
(4) PoI characterization: SURF creates a square region specialists. Our objective when creating a visual dictionary is
centered on the PoI, and oriented along the orientation selected to learn, from a training set of examples, a model that selects
in Step (3). The region is split up regularly into smaller the more representative regions for our problem. The size of
square sub-regions. For each sub-region, the method computes our dictionary must be large enough to distinguish relevant
some simple features at regularly spaced sample points. changes in the images and disregard irrelevant features.
To create the “dictionary”, we need to choose its size , or
Basically, Stages (1) and (2) gives the points of interest (PoIs) number of representative words. During training, the
while Stage (3) assigns the orientation of each PoI and Stage specialists select regions of interest in the analyzed images and
(4) performs the description of each PoI. creating masks for candidate regions more likely to contain the
DR anomaly of interest. The points of interest are then
C. Visual Vocabulary considered in these more likely DR regions. To create the
The creation of the dictionary is outlined in Fig. 1. SURF is dictionary, we can then perform clustering such as k-means for
a good low-level representative feature detector with several finding representative centers for the cloud of PoIs or simply
applications in computer vision. However, SURF-based pick PoIs within the specialist marked regions as we indeed do
approaches are often designed to provide exact matching and in this paper.
they do not translate directly into good results for image A good dictionary is the one that captures properties of the
classification in broad or even constrained domains. Therefore, DR anomaly of interest as well as properties of normal images.
we use the concept of visual vocabularies [12] to capture the Therefore, we create the dictionary using 50 “words”
high-distinctiveness of PoIs while using such discrimination representing the DR anomaly of interest (e.g., bright lesion)
for image classification. and 50 “words” representing normal retinas (non-DR) [12].
In the construction of a visual vocabulary, each region of Note that in the training stage we used common images and
PoIs becomes a visual “word” of a “dictionary”. In the not specific indigenous images.
following, we consider the problem of exudate detection for
D. Training and Classification
the sake of explanation. The approach we discuss in this paper
is general enough to detect other DR-related anomalies as we The visual dictionary was created from the training images,
show in Section IV and V. using the fine selection of candidate regions in normal and
To solve the problem of detecting bright lesions in ocular- abnormal images, which the specialists marked. The selection
fundus images, we select and create a database of training process is only performed in the training stage. Once the
dictionary was created, each of PoIs from the training set was
examples comprising training positive images with exudates
assigned to the closest visual word of the dictionary. This step
and negative images considered normal by specialists. In this
is known as quantization. At the end of the quantization
training stage, we perform the localization of the interest process, a set of feature vectors representing the histogram of
points in all available images using SURF. In this work, we do
5952
the selected visual words for each image are obtained. The Fig. 3 (top) shows an example of a good characterization of
final classification procedure was performed using the Support an aboriginal image in which traces of bright lesions can
Vector Machine (SVM) algorithm. The classifier was trained clearly be seen. This is due to the dominance of the tested
by adding the feature vectors calculated from the training signatures with respected to the typical signature of a DR
images containing positive (e.g., images containing pathology) patient with bright lesions (higher frequencies in positions 1-
and negative (normal images) examples. The parameters for 50).
the SVM were settled during training using SVM grid search.
To analyze multiple pathologies, we can create different
dictionaries (normal vs. bright lesions, normal vs. red lesions)
and train different two-class classifiers that can be combined
later using state-of-the-art machine learning classifier fusion.
5953
test unknown images. For this we selected an operational point candidates using a cross-training methodology that was robust
in which false negatives, for instance, were penalized. against differences in retinal fundus color and therefore easy to
implement worldwide. In addition, the unified proposed
approach allows us to develop different detectors under the
same simple underlying characterization procedure allowing
more than one lesion to be identified at a time therefore
yielding better differentiation in terms of DR disease
progression. As a future direction, we aim at investigating
machine learning feature and classifier fusion techniques in
order to combine different anomaly detectors toward more
discriminative DR vs. Non-DR classifiers.
ACKNOWLEDGMENT
We thank the Federal University of São Paulo medical team
for collecting and grading the ocular-fundus images. In
particular, we express our gratitude to Dr. Eduardo Dib, from
Federal Univ. of São Paulo, for taking the lead in grading the
training images as well as to Dr. Alan Luckie, from the Albury
Eye Clinic, for providing the aboriginal images and initial
classification of the testing retinal images.
REFERENCES
5954